Between Poetry and Precision

Human beings have always had a slightly strange relationship with communication. We spend an extraordinary amount of our lives trying to make ourselves understood, and yet some of the forms of communication we value most seem deliberately designed to resist complete understanding. We want clarity when we are giving directions, writing a contract or proving a theorem, but we are perfectly happy to spend centuries arguing over what a poem means. In one context, ambiguity is a defect. In another, it is the whole point.

Mathematics probably represents one extreme. A mathematical statement is valuable partly because, once the symbols and assumptions have been agreed, there is very little room for personal interpretation. Two people separated by language, culture and centuries should still be able to arrive at essentially the same meaning. Logic works in much the same way. We define terms, state premises and try to remove enough ambiguity that the conclusion follows whether we happen to like it or not.

Poetry sits somewhere near the other extreme. A poem can mean something slightly different to everyone who reads it, and sometimes something different to the same person at different stages of life. A line can remain with us for years precisely because we are not entirely sure why it matters. A painting does not normally come with a specification of the emotional response it is intended to produce, and music would be rather impoverished if every composition had to be accompanied by a paragraph explaining exactly what the listener was supposed to feel.

This difference is interesting because both modes of communication came from the same human mind. We invented mathematics and poetry. We invented formal logic and metaphor. Apparently, we need both the ability to remove ambiguity and the ability to create it.

The things we mean, and the things we leave unsaid

Perhaps this is because human beings do not communicate only information. We communicate moods, intuitions, possibilities, memories, fears and things that we ourselves may not completely understand. Sometimes we want another person to know exactly what we mean. At other times, we want to create a space around what we mean and allow the other person to enter it.

This happens constantly in ordinary conversation. We speak differently to a lawyer than we do to a child, differently to a colleague than to an old friend, and differently again to somebody we love. Some situations reward precision, while others depend heavily on shared context, implication and things that remain unstated. A close friend can sometimes understand an entire paragraph from a raised eyebrow. It is a terribly inefficient communication protocol, but somehow it works.

Humour may be one of the best examples. Any comedian will tell you that if a joke has to be explained, it is more or less dead on arrival. The joke often works because the listener makes the final connection. Timing matters, shared knowledge matters, and sometimes the funniest part is the thing that was never actually said. Once somebody starts explaining the reversal, the hidden assumption or the double meaning, the joke may become more understandable, but it usually becomes much less funny.

There is something particularly amusing, then, about the fact that I now occasionally find myself asking AI why a particular joke was funny. We have somehow reached a point where I can laugh at something, remain slightly uncertain about why I laughed, and then ask a machine to dissect the mechanism for me. The machine can often do a respectable job of it. It can identify the incongruity, explain the reference and tell me where the expectation was subverted. But by the end of the explanation the joke feels a little like a frog in a biology lesson: better understood, perhaps, but not obviously improved by the procedure.

That distinction between understanding and explanation is important. We often behave as if understanding something means being able to make it completely explicit, but human experience does not always work that way. We can understand a piece of music without being able to describe why it moves us. We can feel that a metaphor is exactly right without being able to replace it with a literal sentence. We can understand the sadness in somebody’s voice even when the words themselves contain nothing sad at all.

There are forms of meaning that become weaker when we make them too precise, and that is worth remembering as we begin communicating more and more of our intentions to machines.

A new kind of listener

For most of history, our ability to move between these different styles of communication was exercised mainly with other human beings. Computers were a very different matter. We have, of course, been communicating with computers for decades, but relatively few people had to do it directly, and those who did generally had to learn a language designed for the computer rather than the human.

Programmers became accustomed to an unusually unforgiving listener. A misplaced character, an incorrect type or a badly formed expression could make an instruction mean something completely different, or nothing at all. The computer did not care what we intended. Intention had to be translated into something sufficiently precise for the machine to execute.

Most people were insulated from this. They used buttons, menus and forms, while software developers performed the difficult translation between human intention and machine instruction. The interface was deliberately constrained so that users did not have to think very much about how computers understood things.

Artificial intelligence changes this relationship in a way that I think is much more significant than simply giving us a new generation of software tools. For perhaps the first time, an enormous portion of the population is having complicated conversations with computers in ordinary language. People who have never written a line of code are now asking machines to write, research, analyse, design, summarise, plan, explain and reason. They are not selecting one of five predetermined options from a menu. They are attempting to describe an intention, and describing an intention turns out to be surprisingly difficult.

We say, “Make this better,” and the machine immediately exposes the problem hidden inside that apparently simple sentence. Better in what sense? Shorter? More accurate? More persuasive? More formal? More imaginative? We ask it to design something “clean” or make something “professional”, and discover that those words were carrying a considerable amount of unstated meaning inside our own heads.

This is not really a problem unique to AI. Anybody who has managed people, worked in a team or tried to explain a requirement to a software developer has experienced the same thing. We give what feels like a perfectly clear instruction, receive something completely different from what we imagined, and only then realise that half of the specification never left our head.

AI simply makes this experience unusually frequent. It gives us a listener that is patient enough to receive endless clarification and literal enough to expose our assumptions, while still being intelligent enough to fill in many of the gaps. The result is a strange new pressure on human beings to become more deliberate about how we communicate.

Precision as a new mass skill

People who become effective users of AI quickly start acquiring habits that resemble good engineering communication. They provide context. They state constraints. They give examples. They distinguish between what is required and what is merely preferred. They describe the desired output and, when the result is wrong, they try to identify where their instruction allowed the misunderstanding to occur. None of these are really “prompt engineering” tricks. They are communication skills, and the interesting difference now is scale.

Software developers have been learning this kind of precision for decades because computers demanded it, but programmers were always a relatively small part of the population. Now teachers, lawyers, marketers, accountants, students, managers, designers, doctors and millions of other people are beginning to experience the same pressure. They are communicating with humans during one part of the day and with machines during another, sometimes using the same language but requiring quite different degrees of explicitness. Software developers have been learning this kind of precision for decades because computers demanded it, but programmers were always a relatively small part of the population. Now teachers, lawyers, marketers, accountants, students, managers, designers, doctors and millions of other people are beginning to experience the same pressure. They are communicating with humans during one part of the day and with machines during another, sometimes using the same language but requiring quite different degrees of explicitness.

I suspect this may eventually change the way we communicate with one another as well. If you spend enough time learning to provide context to an AI, separate assumptions from requirements and explain exactly what you are asking for, perhaps some of that discipline follows you back into human conversations. Managers may become better at describing outcomes instead of assuming everybody shares their mental model. Developers may become better at making architectural assumptions explicit. Writers may notice more quickly when an argument depends on something they never actually said.

That would probably be a good thing. Human communication is often remarkably inefficient. We are vague, we forget context, we assume knowledge the other person does not possess, and we routinely confuse what we said with what we intended to say. AI may turn out to be a very effective mirror for this particular weakness.

But there is a danger in learning any useful skill: we may start applying it where it does not belong.

The cost of explaining everything

The modern world already has a strong tendency to reward what can be made explicit. Objectives become metrics, experiences become ratings, conversations become action items and thoughts become bullet points. There are good reasons for all of this. Precision helps organisations function, engineers build things and societies coordinate increasingly complicated systems.

AI adds another incentive. The clearer we are about what we want, the more useful the machine often becomes. There is therefore a very practical reward for removing ambiguity from our language.

What worries me slightly is the possibility that we may eventually confuse this with a general theory of good communication.

Some things should be precise. If I am asking an AI to modify a financial calculation, write a piece of software or summarise a legal document, I would rather not rely on poetic ambiguity. I want assumptions stated and terms defined. But it does not follow that every valuable form of human expression improves as ambiguity decreases.

A poem is not a failed specification. A metaphor is not an inefficient equation. A joke is not incomplete merely because part of its meaning exists in the listener rather than in the words themselves.

In fact, much of what makes human communication beautiful may come from this incompleteness. The listener or reader participates in creating the meaning. We bring our own memory, culture, mood and experience to what is being said. This is why the same poem can mean something different at forty than it did at twenty, and why a song can suddenly acquire significance because of something that happened years after we first heard it.

Perfectly precise communication tries to eliminate that variability. For many purposes, that is exactly what we want. But art often depends on it.

There is also something very human about implication. Sometimes we do not want to state everything directly. Not because we are being dishonest or careless, but because saying less can communicate more. A pause can matter. A badly timed joke can fail even if every word is technically correct. A sentence can be affectionate, threatening, sarcastic or sad without any of those properties appearing in its literal meaning. Human language carries a shadow around the words themselves, and we spend our lives learning how to read it.

I would hate for us to become so impressed by our newfound ability to communicate efficiently with machines that we start treating that shadow as noise.

Learning when not to be precise

Perhaps the real communication skill of the AI age will not be precision itself, but the ability to move consciously between different kinds of precision.

There will be situations in which we need the language of mathematics: definitions, assumptions, constraints, logic and clarity. AI is making that mode of thinking relevant to far more people than before, and I think that is largely a good development. Being forced to explain what we actually mean can expose unclear thinking, hidden assumptions and contradictions that would otherwise remain comfortably unnoticed.

But there will also be times when this is exactly the wrong objective. We will still need stories in which the meaning is not stated, jokes that nobody explains, paintings that refuse to tell us what to think and poems that mean something slightly different every time we return to them.

Human beings have spent thousands of years developing both traditions. We invented mathematics because the world sometimes demands precision, and we invented poetry because the human mind sometimes demands something else. One helps us remove ambiguity from reality so that we can reason about it. The other sometimes reintroduces ambiguity so that we can experience it differently.

AI is now forcing an unprecedented number of people to practise the first of these skills. I suspect it will make many of us clearer thinkers and perhaps better communicators. I hope it does.

But I also hope that, in learning how to make ourselves perfectly understood by machines, we do not begin to believe that being perfectly understood is always the purpose of language. Somewhere between what is said and what is understood lies interpretation, imagination, humour, metaphor and mystery. It is an inefficient space, and no doubt a frustrating one if you are trying to write a specification, but it may also be where much of the beauty of being human lives.